Fix notebook restart flow: clear stale artifacts and guard against empty/corrupt checkpoint CSVs

[charlesmartin14]

Motivation

• The notebook could attempt to resume from existing but empty or corrupted checkpoint CSVs, causing pandas.errors.EmptyDataError and aborting runs.
• When not actually resuming, prior run artifacts should be removed so stale files do not clobber a fresh run.

Description

• Added a remove_stale_run_artifacts() helper and start_fresh logic that clears registry/, per_dataset/, aggregate/, and logs/ when FORCE_RESTART_ALL is True or when restart/resume is disabled (RESTART=False or RESUME_FROM_CHECKPOINT=False).
• Hardened registry cache loading for full_registry.csv with try/except catching pd.errors.EmptyDataError and pd.errors.ParserError, unlinking invalid caches and rebuilding via the scan when needed.
• Hardened sampled-registry reuse for random10_registry.csv with the same guards and fallback resampling when the cached CSV is invalid or empty.
• Made per-dataset checkpoint loading (load_dataset_checkpoint) robust to empty/corrupt metrics CSVs by catching CSV parse errors, removing bad files, and starting the dataset from round 1.

Testing

• Parsed the updated notebook JSON with python -c "import json; json.load(open('notebooks/XGBWW_Random10_LongRun_Alpha_Tracking.ipynb')); print('ok')", which succeeded.
• Verified presence of the new guards/reset messages with grep searches for Starting fresh run, remove_stale_run_artifacts, Invalid cached full registry, Invalid cached sampled registry, and Invalid dataset checkpoint in notebooks/XGBWW_Random10_LongRun_Alpha_Tracking.ipynb, which returned matches.
• Opened and inspected relevant notebook cells to confirm the remove_stale_run_artifacts, safe CSV reads for full_registry.csv/random10_registry.csv, and updated load_dataset_checkpoint behavior were inserted correctly.

---

Codex Task